Phi-3.5-mini-instruct
microsoft/Phi-3.5-mini-instruct · MIT
A 3.8B model that reasons better than models twice its size.
Phi-3.5-mini is trained on carefully filtered, reasoning-heavy data, making it a strong default for math, code and logic on a single consumer GPU.
Score
Setup at a glance
pip install transformers Deep dive
Phi-3.5-mini-instruct is a 3.8B instruction-tuned model built on Microsoft's philosophy of reasoning-dense synthetic data: heavily filtered, textbook-quality training material that lets a small model reason well.
It supports long context (up to 128K tokens), which is unusual at this size, and the model card is refreshingly explicit about limits — primary training in English, known multilingual performance and safety gaps, and documented harm-representation caveats.
The Phi-4 family (including Phi-4-mini and Phi-4-multimodal) is the natural successor to compare against when choosing a small model today.